Although the utilization of mobile technologies has recently emerged in various educational settings, limited research has focused on cognitive load detection in the pen-based learning process. This research conducted two experimental studies to investigate what and how multimodal data can be used to measure and classify learners' real-time cognitive load. The results found that it was a promising method to predict learners' cognitive load by analysing their handwriting, touch gestural and eye-tracking data individually and conjunctively. The machine learning approach used in this research achieved a prediction accuracy of 0.86 area under the receiver operating characteristic curve (AUC) and 0.85/0.86 sensitivity/specificity by only using handwriting data, 0.93 AUC and 0.93/0.94 sensitivity/specificity by only using touch gestural data, and 0.94 AUC and 0.94/0.95 sensitivity/specificity by using both the touch gestural and eye-tracking data. The results can contribute to the optimization of cognitive load and the development of adaptive learning systems for pen-based mobile learning.